Comparative Analysis of Machine Learning Models for Short-Term Distribution System Load Forecasting

Abstract: Accurate electrical load forecasting is crucial for optimizing power system operations, planning, and management. As power systems become increasingly complex, traditional forecasting methods may fail to capture the intricate patterns and dependencies within load data. Machine learning (ML) techniques have emerged as powerful alternatives, offering superior prediction accuracy and the ability to model non-linear and complex temporal relationships. This study presents a comprehensive comparison of prominent ML models: feedforward neural networks, recurrent neural networks, long short-term memory networks, gated recurrent units, and the attention temporal graph convolutional network; for short-term load forecasting of the Energy Corridor distribution system in Houston, Texas. Using a 24-hour look-back window, we train the models on datasets spanning one and five years, to predict the load demand for the next hour and assess performance. Our findings aim to identify the most effective ML approach for accurate load forecasting, contributing to improved grid reliability and system optimization.